An alternative to quartz filters.

[gf]

Once I was a little involved in the analysis of information from acceleration sensors. The system used the AD8552 and AD7767 - 24 bit ADCs. The least significant eight bits were ALWAYS noise without a signal source. That is, I have a slightly biased attitude.

I understand that 10-12-14 bits are applicable. But only with averaging and filtering. These operations require time and computational resources.

But you want to filter anyway. As far as I remember, you did ask for a bandpass filter with quartz filter like characteristics, didn't you?

What I do not understand - what is happening with the noise in the target band?

If we assume a sampling rate of (say) 30MSa/s, then the total noise power is distributed over the 0-15MHz Nyquist band. Assuming a 3kHz target band, only 1/5000 of the total noise power falls into the target band. That's a processing gain of 10*log10(5000)=37dB then. If the ADC's SNR were (say) 80dB in the first place, then you'd get a SNR of even 117dB wrt. the target band's bandwidth.

[radiolistener]

The least significant eight bits were ALWAYS noise without a signal source. That is, I have a slightly biased attitude.

This is normal, there is thermal noise and any amplifier or active RF circuit (include ADC) also adding some power of noise.

Even if you leave ADC input open, it still have internal resistance and that resistance produce thermal noise. So, even open input have noise source

[radiolistener]

But only with averaging and filtering. These operations require time and computational resources. What I do not understand - what is happening with the noise in the target band?

Noise power is distributed across the entire spectrum. When you apply filter, it removes noise power for out-of-band frequencies. So, the total noise power is reduced. In such way noise power is linked with bandwidth.

[SilverSolder]

Dream, meet reality! 

[gf]

Dream, meet reality! 

What exactly do you mean were unreal?

If you look at ADC datasheets, then you see that the white noise assumption is basically granted. Quantizaion noise may not get whitened when the signal and the clock are correlated. But that can be addressed with dithering, so it is not an unsolvable issue either.

One point which still needs to be considered are the other channels, as already pointed out earlier by ejeffrey:

One thing you do need to account for is extra channels.  If you have multiple channels they each can't have a full-scale sine wave simultaneously or you will get clipping. A proper analysis has to take this into account based on your signal.

If all channels are populated, this can have a significant impact. For 2333 channels (=7MHz/3kHz), this costs additional log2(2333)=11.2 bits in the worst case, i.e. if all channels happen to be correlated so that they have their peaks at the same time (which is unlikely, though).

[SilverSolder]

If you operate the ADC at DC, does it stand a better chance of getting closer to its theoretical dynamic range?

I guess an obvious question is -  why have more bits in the ADC than it can actually use?

[gf]

If you operate the ADC at DC, does it stand a better chance of getting closer to its theoretical dynamic range?
I guess an obvious question is -  why have more bits in the ADC than it can actually use?

At full bandwidth these extra bits drown in noise, but for bandwidths significantly lower than fs/2, processing gain can still dig these bits out of the noise floor.

Example: The LTC2217 referenced in a previous post has 16 bits, and a SNR corresponding to about 13 ENOBs. By averaging 64 samples you can drop the noise floor to 1 LSB (of the full 16 bit precision).

[radiolistener]

If you operate the ADC at DC, does it stand a better chance of getting closer to its theoretical dynamic range?

I guess an obvious question is -  why have more bits in the ADC than it can actually use?

to get better dynamic range for DC, you're needs to add a little noise on the input and then apply filtering in digital domain.

But such dithering is not required for radio receiver, because radio spectrum already consists a lot of noise (atmospheric, radio stations, smps interference etc)

[ejeffrey]

  I understand that 10-12-14 bits are applicable. But only with averaging and filtering.

Of course.  But it is impossible to make a 3 kHz bandpass filtering without averaging over a long period of time.

These operations require time and computational resources. What I do not understand - what is happening with the noise in the target band?

Any noise in the target band will pass through the filter exactly like the signal and that doesn't matter whether it is a crystal filter, analog LC filter, digital filter, or mixer based filter. 

[MikeP]

 So. First of all, I want to express my gratitude and thanks to Mike - mawyatt. Filter experiments have become a New Year's adventure. You can't buy it in travel agencies!

 We are talking about a modest device - a ten-bit counter, small buffers and 7002 transistors. To my surprise, the counter is operational up to 80 MHz. Input resistor - 5kOhm, capacitors - 47nF.

  Yes, it's a filter. But NO, it doesn't look like a quartz filter. Moreover, there are two big disadvantages. The center frequency of the filter DOES NOT coincide with the LO frequency. I find it hard to find an explanation.
  The second phenomenon is the presence of a parasitic signal at the output when there is no signal at the input. I think there are several reasons: different pulses lengths, different rise and fall speed in individual cells. Some kind of leaks. That is, the capacitors always have some offset, which leads to the generation of LO/10.

 I must say that I have read a dozen academic papers on this subject. The shortcomings I found were not mentioned. There are no examples of frequency response. Now everyone has a little understanding of the circuit. Thanks.

[tggzzz]

When I built such a filter in ~1979 (IIRC using 4066 analogue switches), I also observed such a waveform.